Occasionally situations arise in which a measurement does not lend itself to such traditional methods of reliability estimation as the test-retest, parallel-test, or internal consistency methods, for example, because a single item variable or an index based on heterogeneous data is involved. In this paper, it is proposed to base reliability estimation in such situations on estimates of validity coefficients as lower bounds. These lower bounds can be maximized by a deliberate selection of predictor variables, both in the case of single item variables and heterogeneous indices. Two selection procedures are examined and compared, one based on expert judgment and one on backward deletion of predictors with cross validation as a provision against chance capitalization. Some examples presented in the paper suggest that these methods provide satisfying estimates.